# The GTM Harness: State of the Art, August 2026 > [!abstract] How to read this document > **Part A** is my brief, written by me. **Part B** is the questions I need answered. **Part C** is AI-generated research — well sourced, but not settled fact. The disclosure at the end explains exactly how it was produced and where it is weak. > Start with A and B. Read C as a map of the territory, not as a decision. --- # Part A — The brief %% Written by Karan. Everything in this section is human-authored. %% ## What I asked for I want to build a really good go-to-market harness that sits across the modular offering we give fellows — the seven pillars in the partnership doc: Growth Strategy, BD & Partnerships, Sales Enablement, Customer Success, Marketing & Comms, GTM Engineering, IP Commercialisation. It needs skills and agents that each do different things, and I need the right technical harness to bring it all together. So I asked for research on the best open-source go-to-market harnesses out there, so I understand the state of the art. I know Rox and Clay are very good — I wanted to know what the people who do this well actually use, and what exists in the open. ## The constraints that matter to me > [!important] These are the evaluation criteria, not nice-to-haves > Judge everything in Part C against these five. If a recommendation fails one of them, say so. 1. **It has to be adoptable by anybody.** Not just us, not just our fellows. Someone should be able to pick it up and use it in their own context. 2. **It has to be CRM-agnostic.** We use Attio. The next person uses HubSpot, or Salesforce, or nothing yet. ==Swapping the CRM should not mean rebuilding.== 3. **It has to be agent-agnostic.** Different intelligence agents, different models, different runtimes. We should not be locked to one vendor's harness. 4. **It has to be genuinely modular.** Take one pillar, use it on its own. Take three, compose them. Nothing should require adopting the whole thing. 5. **It has to be super easy to use.** If it needs a GTM engineer on $135k to operate it, we have built the wrong thing. ==Our fellows are founders, not RevOps.== ## What I want you to do Read Part C and pressure-test it against those five constraints. I want your honest assessment, not a summary. Come back to me with answers to the questions in Part B. --- # Part B — Critical questions I need answered > [!question] Ground rule > Where Part C doesn't answer one of these, **say so plainly**. A clear "this isn't answered" is more useful to me than a confident guess. ## On the core choice > [!danger] These two questions attack the research's own recommendations. That is deliberate. 1. **Is "there is no open-source GTM harness, so build the orchestration layer yourself" actually the right conclusion — or did the research miss something?** Go looking independently. If a credible open-source GTM harness exists, I want to know about it. 2. **The research recommends the Claude Agent SDK as the core, and that is not open source** — it is Anthropic commercial licensing. That conflicts head-on with constraint 3. Is the proposed hedge — keeping all logic in portable `SKILL.md` files and MCP servers — actually sufficient? What would we genuinely lose if we had to move to a different runtime in a year? 3. **Same question for the CRM.** The research leans on Attio's MCP server, which conflicts with constraint 2. What does a real CRM-agnostic adapter layer look like, and what does it cost us in capability to build one? ## On modularity and ease of use 4. **What is the smallest useful unit someone can adopt?** One skill? One pillar? Describe the "hello world" — what does a founder do in their first 30 minutes, and what do they get out of it? 5. **What does the setup actually require?** Be concrete: accounts, API keys, command line, cost. If the honest answer is "a technical person for a day," say that. 6. **Which of the seven pillars can be delivered as pure skills with no infrastructure at all, and which genuinely need a running system?** I suspect the split is larger than it looks, and it matters a lot for how we ship this. ## On the hard constraints 7. **Part C says two layers can't be replicated in open source — contact data and email deliverability.** Do you agree? What is the cheapest honest path through each for a seed-stage founder? 8. **What breaks first at scale?** If twenty fellows run this simultaneously, what falls over — cost, rate limits, isolation, our support burden? 9. **Where is the real security exposure?** Part C flags prompt injection in third-party skills. What is the minimum safe setup for a system that can write to a CRM *and* send email? ## On the decision 10. **What would you actually build first, and what would you deliberately not build?** I would rather ship one pillar that works than seven that half-work. 11. **What is the strongest argument against this whole approach?** Steelman "just pay for Clay and Attio and skip the harness." Then tell me why you do or don't buy it. > [!tip] How to come back to me > Short written answers, in your words, with the sources you checked. Long AI-generated writeups are not what I'm after here. --- # Part C — The research > [!warning] AI-generated > This section was produced by AI. Every claim is linked to a source; vendor-authored and self-reported figures are flagged inline. See [[#Disclosure how Part C was produced]] at the end. **Purpose:** Understand the best open-source and open-API building blocks for a modular GTM harness that can deliver the seven-pillar partnership offering across a portfolio of fellows. ## 0. The headline finding, stated plainly > [!important] There is no open-source Clay, and there is not going to be one > Nor is there an open-source Rox. What exists is a mature, high-quality open-source stack for **everything except two layers** — proprietary contact data, and email inbox reputation. Those two are bought, not built, and every serious player buys them. That reframes the question. The right target is not "an open-source Clay." It is a **Cargo-shaped orchestration runtime**: a modular harness holding skills, agents, memory, policy gates and evals, sitting on thin adapter interfaces to rented data and rented sending, with a CRM as system of record. The commercially interesting fact underneath this: in March 2026 Clay split its pricing into **two meters — Data Credits and Actions**. Actions meter *execution steps* (HTTP calls, CRM syncs, AI tasks) ==even when you bring your own API keys==. Clay explicitly closed the "just use my own vendor contracts" escape hatch, because orchestration is where the margin now lives. That orchestration meter — roughly $450/mo at Growth tier, $5k+/mo from the AI SDR vendors — is a DAG engine with retries. **That is the layer you can own.** ## 1. What the state of the art actually is ### Clay — the reference implementation *$150M ARR; $5B employee tender, January 2026.* Clay is a **spreadsheet runtime for GTM data**. Tables of people and companies; columns that are either fields or *actions* (enrichment call, HTTP request, AI prompt, CRM write). Its own four-layer model: Data → Orchestration → Execution → Agents. What matters architecturally: - **Waterfall enrichment across 200+ providers.** The core value. No single vendor exceeds ~30–40% coverage; a well-ordered cascade reaches 70%+. - **Claygent** — a web-research agent returning structured multi-column output. Model lineup now includes Clay-native Helium/Argon plus OpenAI, Anthropic, Gemini — *and open-weight models (Kimi K2.6, GLM 5.2)*. Clay is commoditising its own AI layer. ==The agent layer is not the moat.== - **Signals Hub** — career movement, web intent, social listening, news, first-party product usage, tech-stack change; plus "turn any enrichment or AI query into a signal." - **Functions** — reusable GTM logic defined once, reused across tables. (Note: "Claybooks" are case-study recipes, not a runtime. Functions are the executable unit.) - **Clay MCP server** — OAuth; works in Claude, ChatGPT and Codex; plus a CLI with headless device login. Critically, `run_subroutine` **executes your workspace's custom Functions from an LLM client** — a remote execution surface, not just a data read. Capped at 100 results per search, so not a bulk pipe. > [!bug] Limits worth knowing > Clay does **not verify emails itself** — quality depends entirely on which waterfall provider returned the hit, so serious stacks bolt a verifier on afterwards. Credit burn is nonlinear: a 5-provider waterfall runs ~15 credits/contact. HTTP integrations, CRM auto-sync, Web Intent and webhooks are gated to Growth+ ($446/mo annual). > Structurally, Clay is a *table* abstraction — long-running stateful per-account agents fit it badly. That gap is exactly what Rox, Unify and Cargo attack. ### Rox — the agent-swarm thesis *$1.2B valuation on ~$8M ARR.* Founded 2024 by Ishan Mukherjee (ex-New Relic CGO) and **Chris Ré** (Stanford). Sequoia seed; General Catalyst + GV Series A. The differentiated idea: **an always-on swarm of agents assigned per account**, each specialised — news monitoring, sentiment, opportunity identification. And it is **warehouse-native**: it sits *on top of* Salesforce, grounding a knowledge graph natively in Snowflake using zero-copy sharing, so raw data never leaves customer control. Tiered model routing — cheap models at the data layer, mid-tier for reasoning, frontier plus realtime for interaction. **Rox builds no proprietary models.** > [!note] The distinction that matters > Clay owns *acquisition of unknown data*. Rox owns *reasoning over data you already have*. > Rox's job is far more tractable to replicate in open source, because there is no proprietary dataset involved — it's a warehouse, a knowledge graph and a durable agent scheduler. ### The rest of the field, compressed | Player | The one thing to take from them | |---|---| | **Unify** ($40M Series B, Battery + OpenAI Fund) | Sells **managed deliverability** — mailbox creation, warmup, rotation. That tells you how miserable that layer is. | | **Common Room** | **Person360** identity resolution + official MCP server **and a CLI for headless pipelines**. Most agent-friendly commercial signal layer. | | **Apollo** (acquired Pocus, Mar 2026) | Official MCP, OAuth, **50+ actions including writes** (sequences, contacts, deals). Best single-credential substitute for a big chunk of Clay's surface. Search/enrich burn credits; sequence and CRM ops don't. | | **Attio** — *our CRM* | 30,000 customers; 400M API calls/week; **2.6M MCP calls/month**. REST + webhooks + App SDK on **every tier including Free**. MCP server and App SDK need Pro ($99/user/mo). MCP built purpose-first, not as an API wrapper: TOON encoding for token efficiency, fewer/richer tools, dynamic client registration so the agent authenticates *as the user* and Attio's permission model applies. **Sequences and Workflows now ship on all plans.** | | **Octave** | The "GTM brain" — positioning, ICP and messaging as structured living context consumed by other tools, via REST + MCP. Directly relevant to the Growth Strategy pillar. | | **Cargo** | **The most useful reference architecture published.** "Centralise logic, distribute execution." Deployment mirrors software release: shadow mode → canary → policy hardening → human-in-the-loop gates → graduated autonomy. Metrics: revenue latency, routing precision, data health, rep leverage. | | **11x** | Cautionary tale. $20M Benchmark Series A, then 2025 allegations of inflated ARR and demo-vs-production gaps. | > [!tip] The pattern that matters > The **infrastructure** players (Clay, Attio, Apollo, Common Room, Octave) all ship MCP. The **outcome** players (11x, Artisan, Landbase, Warmly) don't — the product *is* the closed workflow. > Our harness composes trivially with the first group and not at all with the second. Which is fine: the second group is the layer we're replacing. ### The role, in numbers From the [2026 State of GTM Engineering survey](https://knowledge.gtmstrategist.com/p/the-2026-state-of-gtm-engineering) (n=228): Clay at 84% adoption (96% among agencies), CRM 88%, **AI coding tools ~70%**. US median base $135k; coding ability is worth a $40–45k premium. Only 45% of orgs clearly understand what a GTM engineer does. 26% name integrations and closed ecosystems as their top frustration. ## 2. The open-source harness layer This is where the genuinely good open source is. ### The orchestration core | Option | License | Verdict for a GTM harness | |---|---|---| | **Claude Agent SDK** | Anthropic Commercial ToS — **not OSI open source** | The Claude Code agent loop as a library: hooks, subagents, MCP, permissions, sessions, auto-compaction, and it **loads `.claude/` skills, agents and memory straight off disk**. Nothing in the open field matches it for a skills-first design. It is also the one closed, load-bearing component. | | **LangGraph** (38.1k) | MIT | Best open answer for the *deterministic* path: durable checkpointing, human-in-the-loop interrupts, time travel. Use it where money moves. LangSmith (the good observability) is proprietary. | | **Mastra** (27.1k) | Apache-2.0 core + `ee/` carve-out | Strongest TypeScript agent framework — agents, workflows, RAG, evals, memory, MCP, local playground. Younger API, breaking changes. | | **Temporal** (21.7k) | MIT | Gold standard for durable execution. A 3-week outbound sequence with retries, timers and approval gates *is* a Temporal workflow. Heavy ops surface. | | **Trigger.dev** (16k) | Apache-2.0 | The better permissive pick for TS durable background jobs — retries, queues, **concurrency keys** (useful for per-tenant rate limits), self-hostable. | | **n8n** (198k) | **Sustainable Use License — not OSI** | 1,500+ connectors is the real reason to have it. **The SUL restricts offering n8n itself as a service to third parties** — a genuine problem if we host one instance per fellow. | | **Windmill** (17.5k) | AGPLv3 + Apache-2.0 dual | Safer than n8n if reselling. Rust engine, nsjail sandboxing, workspaces as a tenancy boundary. | | **Pydantic AI** (17.5k) | MIT | Type-safe structured output (lead scores, ICP fits, CRM writes). Very active. | | **Letta** (24.3k) | Apache-2.0 | Agents as stateful services with self-editing memory. Supports Agent Skills. Opinionated. | | **CrewAI** (57.2k) | MIT | Fast to prototype role-playing crews. Abstraction taxes; weak durability. | > [!failure] Do not start here in 2026 > **AutoGen** (officially maintenance mode) · **OpenAI Swarm** (dead, superseded) · **Dify/Flowise** as the *core* (weak git/version story, licence friction) · **Inngest server** (SSPL) · **Zep self-hosted** (community edition deprecated). ### The single most important standard: Agent Skills > [!success] This is the piece that makes the offering portable > Agent Skills is now a **governed open standard**, not an Anthropic feature — published December 2025 and stewarded by the **Linux Foundation Agentic AI Foundation**, the same home MCP joined. A skill is a directory: `SKILL.md` (YAML frontmatter + markdown) plus optional `scripts/`, `references/`, `assets/`. Progressive disclosure — ~100 tokens of metadata at startup, full `SKILL.md` on activation, referenced files on demand. **~40 products already implement it:** Claude/Claude Code/Cowork, ChatGPT + Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI, JetBrains Junie, Amp, Kiro, OpenCode, OpenHands, Goose, Letta, Spring AI, Databricks, Snowflake Cortex, Pulumi. **Why this matters for us:** every GTM play in the offering — ICP qualification, account research, sequence writing, objection handling, the LinkedIn Audit agent, BioCraft — becomes a markdown file in a git repo. Diffable. Reviewable by PR. Portable to a different harness. ==Our IP becomes the skill library, not the runtime.== > [!danger] Security, and this is load-bearing > A 2026 ecosystem audit found **36% of tested public skills contain prompt-injection vulnerabilities**, averaging 6.3 issues per skill. (Vendor-published — treat the magnitude as directional, but the direction is right.) > **Never install third-party skills into a harness with send-email or write-to-CRM permissions.** Vendor them, review them, pin them to commit SHAs. ### Distribution: plugins and marketplaces Claude Code **plugins** fit our model exactly. One directory bundles `skills/`, `agents/`, `hooks/hooks.json`, `.mcp.json` and a `settings.json`. **A marketplace is just a git repo** with `.claude-plugin/marketplace.json` — *including a private repo*. Skills are namespaced `/plugin-name:skill-name`, so a fellow's local overrides coexist with the shared library. This is how "proprietary GTM skills" ship as a versioned product rather than a Notion doc. ### Memory and context The underrated answer is **file-over-app**: ICP, positioning, battlecards, objection handling and per-account dossiers live as **markdown in a git repo** — the same substrate as `SKILL.md`. Diffable review, PR-based approval of what the agent believes, no embedding drift, and the agent reads with grep instead of a similarity search that returns the wrong chunk. Add **pgvector in Postgres** for transcripts and email history — accounts and deals are already relational, and one Postgres per tenant is a free isolation boundary. Add **Mem0** (62.8k, Apache-2.0) only for cross-session per-contact facts, and only once the pain is felt. **Skip Zep** — community edition deprecated, effectively cloud-only. ### Evals and observability - **Langfuse** (32.4k, MIT except `ee/`) — trace/eval backbone, self-hostable, one project per tenant, prompt versioning we'll want for outbound copy. - **promptfoo** (24.2k, MIT) — CI gate, and specifically its **red-team mode**. A GTM agent with send permissions is a prompt-injection target via inbound email and scraped web pages. ==Not optional.== - **DeepEval** (17.5k, Apache-2.0) if the team thinks in pytest. - **Not options:** Braintrust core is proprietary; LangSmith is proprietary SaaS. > [!warning] The trade-off across all of them > Open eval tools score *offline* quality well (did the email match the brief?) and attribute *outcomes* badly (did it book meetings?). Wire reply-rate and meetings-booked back from the CRM as the real signal; treat LLM-judge scores as a regression guard only. ### Multi-tenant isolation across a portfolio 1. **One git repo per fellow** — `gtm-<company>` containing `.claude/` (skills, agents, hooks, settings), the markdown vault (ICP, positioning, battlecards, dossiers), and `.mcp.json` with that company's credentials. Config, memory *and* audit log in one artifact. 2. **One shared private plugin marketplace** — the common GTM plugin. Fellows pin versions; we fix once and roll out. 3. **One ephemeral container per run** — credentials injected at start, egress allowlisted to that fellow's CRM and email domains only. Anthropic's example devcontainer ships a **default-deny iptables firewall**; that's the standard pattern. 4. **Separate credentials and separate Postgres per fellow.** An injected prompt in one company's inbound email must not be able to reach another company's CRM. 5. **git worktrees** are for parallel agents *within* one tenant, not across tenants — they share a filesystem and credentials. > [!note] Enforceability caveat > Only the built-in Bash sandbox is enforceable by Claude Code itself. Devcontainers are a convention, not a boundary. If isolation must be a hard control, it has to be container/VM plus our own CI runner policy. ## 3. The data, CRM and execution layer ### CRM **Keep Attio.** Its MCP server is genuinely well-built — purpose-designed rather than an API wrapper, TOON encoding, and the agent authenticates as the user so permissions apply. Sequences and Workflows now ship on all plans, removing most of the need for a separate sequencer. > [!info] Budget note > API and webhooks are on Free, but **MCP and the App SDK require Pro at $99/user/mo**. If we ever need a self-hosted mirror or OSS fallback: **Twenty** (48.9k, AGPL-3.0 core with a proprietary enterprise carve-out) is the only credible one — v2.0.0 in April 2026 shipped model-agnostic agents in workflows and a **native MCP server**. Everything else — EspoCRM, SuiteCRM, Odoo, Erxes, Krayin — is legacy or agent-hostile. **Atomic CRM** (MIT, React + Supabase) is the underrated option if we'd rather own the Postgres schema outright. ### Research and enrichment - **Crawl4AI** (68k, **Apache-2.0**) over Firecrawl (AGPL-3.0). Same job, permissive licence, no API keys self-hosted, LLM-ready markdown, stealth mode, Docker + JWT gateway. ==Take the Apache one.== - **Search APIs are required infrastructure, not a convenience.** SearXNG is a metasearch *proxy*, not an API — it gets rate-limited under agent load. Budget: Exa ~$0.005/search, Tavily ~$0.008, DataForSEO ~$0.002, Serper $1/1,000. - **Browser automation:** Stagehand (MIT, Browserbase) is more production-grade than browser-use — self-healing selectors, act/extract/observe. The paid layer is proxies, CAPTCHA and session persistence. > [!bug] Company and people data — the reality > - **Companies House API** — genuinely free, UK-only, officers/filings/SIC codes. Best free signal source going. > - **OpenCorporates is no longer meaningfully open** — £2,250/yr for 500 calls/mo, annual only, failed searches still burn quota. > - **Crunchbase killed its free tier**; Pro $99/mo for full API. > - **People Data Labs** ~$0.25–0.28/person credit, ~$0.065–0.10/company. > - **Proxycurl is dead.** LinkedIn sued January 2025; Proxycurl shut down 4 July 2025 at ~$10M ARR rather than fight. ==Treat every LinkedIn-derived data vendor as carrying existential continuity risk.== **On building our own waterfall:** the code is a few hundred lines — an ordered provider chain with per-field confidence and cost caps. The moat in Clay is the vendor contracts and credit arbitrage, not the logic. Measured provider characteristics: Hunter 89–96% valid but low coverage; Prospeo 81–94%, best on US enterprise; Findymail 60–67% valid, high coverage; **Wiza 67–85% coverage but only 15–55% validity — actively dangerous to sender reputation**; LeadMagic consistently under 40%. Budget **$0.08–0.12 per valid email** using a multi-provider combo. ### Signals - **JobSpy** (MIT) — job postings across Indeed, Google, LinkedIn, Glassdoor, ZipRecruiter. From its own README: Indeed is best and unthrottled; LinkedIn rate-limits around page 10 and *"proxies are a must basically."* Use Indeed + Google as primary. - **Miniflux** (Apache-2.0, Go single binary, clean REST API) for RSS/news into a pipeline. - **GDELT** — free global news event/tone/entity extraction. Genuinely under-used in GTM. - **HN:** Algolia's HN Search API, free, no key. **GitHub:** GraphQL + webhooks, free. **Reddit:** the weak link — official API now paid above free tier, terms hostile to commercial use. - **Visitor identification: no OSS exists and none can.** It depends on proprietary identity graphs built from data co-ops, and person-level de-anonymisation is **not GDPR-viable in the EU/UK** anyway. The OSS-adjacent ceiling is IP→company firmographics (IPinfo, MaxMind, self-built ASN→org) — *account* level, never person level. - **PostHog** (MIT core + proprietary `ee/`) for product-usage signal — but its own README caps self-host at **~100k events/month with no support**. In practice a paid product. ### Sending and sequencing > [!danger] This is the second layer we cannot win, and the reason is regulatory, not technical > 2026 bulk-sender rules bite at **5,000+ emails/day** to consumer domains: SPF *and* DKIM passing with at least one aligned to the From domain, DMARC published, FCrDNS on sending IPs, TLS, RFC 8058 one-click unsubscribe honoured within 2 days, spam complaints under 0.1%. > Enforcement escalated hard: **Microsoft began rejecting non-compliant mail in May 2025; Gmail moved from temporary 421 deferrals to permanent 550 rejections in November 2025.** Postfix and **Postal** (16.7k, MIT) send mail fine. What's hard is: mailbox supply (dozens of real aged Workspace/M365 mailboxes on secondary domains — a procurement problem); **warmup networks**, whose value is a large pool of cooperating real mailboxes, i.e. a network effect a small private pool makes *worse* than useless; reputation as a slowly-accrued asset destroyable in a day; and empirically-tuned inbox rotation and volume ramping. **Rent it.** Smartlead Pro is $94/mo with API. Attio Sequences are included on all plans. **Resend** ($20/mo Pro, 50k emails) is the pragmatic default for transactional and newsletter. Wrap whichever we choose behind an interface so it's swappable. **There is no credible open-source sequencer.** Nothing OSS does inbox rotation + warmup + reply/bounce detection + branching cadences. Options: Attio Sequences (cleanest, already paid for), or build cadence as agent state in our own Postgres and send via Resend — arguably *better* for an agent harness, since the agent then owns the branching logic. > [!failure] WhatsApp > Only Meta's Cloud API or a BSP (360dialog, Twilio, Infobip). Since 1 July 2025 billing is per-message not per-conversation; each delivered template is charged individually (~$0.025 US marketing). Customer-initiated service conversations are free. > **OSS libraries (Baileys, whatsapp-web.js) reverse-engineer WhatsApp Web, violate ToS, and get numbers banned. Do not build the Customer Success pillar on them.** **Inbound support:** **Chatwoot** (33.9k, MIT core + proprietary `ee/`) is the clear winner — omnichannel including WhatsApp, ships its own AI agent ("Captain"). Papercups is in maintenance mode; Typebot is Functional Source License, not OSI. Nothing OSS approaches Sierra; Chatwoot plus our own Claude agent on its API is the realistic build. ### Content, brand and marcomms > [!success] This is where open source is quietly excellent and nobody looks > Design-system-as-code fully covers our "bespoke design system per company" commitment. - **Penpot** (58.5k, MPL-2.0) — **native design tokens as a first-class feature and open file formats**. The only design tool where an agent can legitimately read and write source files. - **Tokens Studio for Figma** (MIT) — tokens as JSON with GitHub sync. - **Style Dictionary** (Apache-2.0, Amazon) — tokens JSON → CSS/iOS/Android/JS. The build-layer standard. **The pipeline:** Penpot or Tokens Studio → tokens JSON in git → Style Dictionary → CSS vars / Tailwind config → agent renders HTML/SVG → headless Chromium screenshot for raster. Fully OSS, fully deterministic, no Canva or Figma API dependency. ==Every fellow gets a tokens file in their repo, and every asset the harness generates is on-brand by construction.== - **Publishing/newsletter:** **Ghost** (MIT) — Content API + Admin API + webhooks, members and email newsletters built in. Listmonk (AGPL) if we want list management decoupled. - **SEO/AEO:** thin but emerging — `geo-optimizer-skill` (AEO/GEO audit and citation tracking across ChatGPT/Perplexity/Gemini/AI Overviews, with an MCP server), `awesome-aeo-seo-tools`. Classic technical SEO is well covered. **But keyword volume, SERP position and backlink graphs have no OSS equivalent** — DataForSEO at $0.002/search is the cheapest programmatic route. > [!warning] LinkedIn — the hard truth for the Marcomms and LinkedIn Audit pillars > Free self-serve scopes are Sign In (OIDC), `w_member_social` posting, and Verified. **Sales Navigator API closed to new applicants in 2026.** Marketing Developer Platform runs ~$699+/mo with 4-week-to-4-month approval and unexplained rejections. Blocked for almost everyone: messaging, connection lists, people/company search, detailed profile data. > hiQ v. LinkedIn does not help — it addressed CFAA criminality, not contractual permission. > ==Design for: agent drafts → human posts.== Our existing LinkedIn Audit agent working from an uploaded profile PDF is exactly the right pattern, and we should keep it that way deliberately. ## 4. Recommended architecture ```mermaid graph TD V["CONTEXT VAULT (git, markdown)<br/>ICP · positioning · battlecards<br/>brand tokens · account dossiers"] H["HARNESS<br/>Agent SDK + Agent Skills (SKILL.md)<br/>shipped as a private plugin marketplace<br/>skills · subagents · hooks · MCP client"] S["SIGNAL<br/>JobSpy · Miniflux · GDELT<br/>GitHub · PostHog · Crawl4AI"] D["DATA<br/>waterfall skill →<br/>PDL · Hunter · Findymail<br/>Companies House"] X["DECISION<br/>LangGraph<br/>deterministic send path"] A["ACTIVATION<br/>Attio Sequences · Resend<br/>Chatwoot · WhatsApp API<br/>Ghost · Penpot"] R["SYSTEM OF RECORD<br/>CRM via MCP<br/>+ Postgres + pgvector"] F["DURABILITY Trigger.dev / Temporal<br/>OBSERVABILITY Langfuse + promptfoo<br/>ISOLATION container per run · per-tenant creds"] V --> H H --> S H --> D H --> X H --> A H --> R S --> F D --> F X --> F A --> F R --> F ``` > [!important] Three rules that make this work > 1. **Centralise logic, distribute execution.** ICP, scoring, enrichment order, routing rules and playbooks live in one place (the vault plus skills). Execution fans out to the CRM, Slack, browser, sequences. > 2. **Research can be agentic; the send path cannot.** Approve → personalise → send → log is a state machine with checkpoints, not a model deciding. > 3. **Graduated autonomy, like a software release.** Shadow mode → canary segment → policy hardening → human-in-the-loop gates → autonomy. Never ship an agent straight to send. ## 5. How this maps to the seven pillars | Pillar | Harness component | Build / Buy | |---|---|---| | **1. Growth Strategy** | ICP and positioning as structured markdown in the vault; `icp-qualification` and `market-mapping` skills. Octave is the commercial proof this layer has value. | **Build** — this is the differentiated IP | | **2. BD & Partnerships** | Signal layer (JobSpy, GDELT, Miniflux, GitHub) → account scoring skill → CRM. Intro-graph mapping over our own network data. | **Build on OSS** | | **3. Sales Enablement** | Penpot/Tokens Studio → Style Dictionary → per-fellow brand tokens in their repo. One-pagers, customer decks, MOUs, pilot agreements as skills with `assets/` templates that render on-brand by construction. | **Build — fully OSS, a clean win** | | **4. Customer Success** | Chatwoot (MIT) plus a Claude agent on its API. WhatsApp via Meta Cloud API only — never Baileys. | **Buy the channel, build the agent** | | **5. Marketing & Comms** | Ghost for site/newsletter. `geo-optimizer-skill` for AEO. LinkedIn Audit and BioCraft stay as PDF/paste-driven skills — ==that's the correct design given LinkedIn's API reality, not a limitation.== | **Build on OSS; DataForSEO for SERP** | | **6. GTM Engineering** | The harness itself: agent runtime + Agent Skills + private plugin marketplace + per-fellow config repo + CRM MCP. Langfuse and promptfoo as the quality gate. | **This is the product** | | **7. IP Commercialisation** | The skills library *is* the IP. Because Agent Skills is an LF-governed open standard with ~40 implementations, the library is portable and licensable independent of any runtime. Worth taking seriously as a patent/licensing surface. | **Build — and note the strategic point** | ## 6. Suggested next steps > [!todo] In order > - [ ] **Decide the runtime explicitly.** The Agent SDK gives the best harness and is the one closed component. Mitigate by keeping 100% of GTM logic in `SKILL.md`, MCP servers and markdown — portable to OpenCode, Goose or Letta Code, which already read the same skills. **Don't skip making the decision consciously.** *(See Part B, Q2.)* > - [ ] **Stand up one fellow end-to-end** before generalising. One repo, one CRM workspace, three skills, one signal source, one send path with a human gate. > - [ ] **Write the waterfall as a skill early.** A few hundred lines, and the thing that most obviously replaces Clay spend. Start with Companies House (free) plus one paid person-data provider. > - [ ] **Put promptfoo red-teaming in CI on day one.** The harness will read inbound email and scraped web pages while holding CRM write and email send permissions. That combination is the attack surface. > - [ ] **Check the licences before they become architecture.** n8n's Sustainable Use License and Dify's modified licence both restrict offering the product as a service to third parties. > - [ ] **Budget the two rented layers:** data at $0.08–0.12 per valid email; sending at $20–95/mo per fellow; search at $0.002–0.008/query; CRM MCP access at ~$99/user/mo. Everything else is free or self-hosted. ## Sources **Commercial landscape:** [Clay pricing](https://www.clay.com/pricing) · [Clay $5B tender](https://www.businesswire.com/news/home/20260128514638/en/Clay-Announces-Second-Employee-Tender-Offer-in-Nine-Months-at-a-$5B-Valuation) · [Sacra: Clay](https://sacra.com/c/clay/) · [Clay AI pricing docs](https://university.clay.com/docs/ai-pricing) · [Clay MCP](https://www.clay.com/mcp) · [Clay GTM Alpha](https://www.clay.com/blog/gtm-alpha) · [Sequoia on Rox](https://sequoiacap.com/article/partnering-with-rox-every-seller-needs-an-agent-swarm) · [OpenAI: Rox](https://openai.com/index/rox/) · [Rox on Snowflake](https://medium.com/snowflake/rox-the-next-generation-revenue-operating-system-powered-by-snowflake-08b5c6a9a963) · [Unify Series B](https://www.unifygtm.com/blog/series-b) · [Common Room MCP & CLI](https://www.commonroom.io/product/mcp-cli/) · [Apollo MCP docs](https://docs.apollo.io/docs/apollo-mcp) · [Attio MCP engineering post](https://attio.com/engineering/blog/building-the-attio-mcp-server) · [Attio pricing](https://attio.com/pricing) · [Cargo GTM engineering playbook](https://www.getcargo.ai/blog/gtm-engineering-playbook-2026-autonomous-workflows) · [a16z: Death of a Salesforce](https://a16z.com/ai-transforms-sales/) · [2026 State of GTM Engineering](https://knowledge.gtmstrategist.com/p/the-2026-state-of-gtm-engineering) **Harness layer:** [Claude Agent SDK](https://code.claude.com/docs/en/agent-sdk/overview) · [Claude Code plugins](https://code.claude.com/docs/en/plugins) · [Sandbox environments](https://code.claude.com/docs/en/sandbox-environments) · [Agent Skills spec](https://agentskills.io/specification) · [agentskills/agentskills](https://github.com/agentskills/agentskills) · [MCP joins AAIF](https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/) · [Agent Skills ecosystem report 2026](https://agentman.ai/blog/agent-skills-ecosystem-report-2026) · [LangGraph](https://github.com/langchain-ai/langgraph) · [Mastra](https://github.com/mastra-ai/mastra) · [Temporal](https://github.com/temporalio/temporal) · [Trigger.dev](https://github.com/triggerdotdev/trigger.dev) · [Windmill](https://github.com/windmill-labs/windmill) · [n8n](https://github.com/n8n-io/n8n) · [Langfuse](https://github.com/langfuse/langfuse) · [promptfoo](https://github.com/promptfoo/promptfoo) · [Mem0](https://github.com/mem0ai/mem0) · [file over app](https://stephango.com/vault) **Data & execution:** [Twenty](https://github.com/twentyhq/twenty) · [Atomic CRM](https://github.com/marmelab/atomic-crm) · [Crawl4AI](https://github.com/unclecode/crawl4ai) · [Stagehand](https://github.com/browserbase/stagehand) · [JobSpy](https://github.com/speedyapply/JobSpy) · [PostHog](https://github.com/PostHog/posthog) · [Postal](https://github.com/postalserver/postal) · [Listmonk](https://github.com/knadh/listmonk) · [Resend pricing](https://resend.com/pricing) · [Chatwoot](https://github.com/chatwoot/chatwoot) · [Penpot](https://github.com/penpot/penpot) · [Style Dictionary](https://github.com/amzn/style-dictionary) · [Tokens Studio](https://github.com/tokens-studio/figma-plugin) · [Proxycurl shutdown](https://nubela.co/blog/goodbye-proxycurl/) · [LinkedIn API 2026 guide](https://connectsafely.ai/articles/linkedin-api-complete-guide-2026) · [PDL pricing](https://support.peopledatalabs.com/hc/en-us/articles/25794271805211-Pricing-credits) · [Email enrichment benchmark](https://www.growthtoday.co/blog/best-email-enrichment-providers) · [Red Sift bulk sender requirements](https://redsift.com/guides/bulk-email-sender-requirements) · [WhatsApp API pricing 2026](https://blueticks.co/blog/whatsapp-business-api-pricing-2026) --- # Disclosure: how Part C was produced > [!warning] Part C of this document was generated by AI > Parts A and B are Karan's own brief and questions. - **Produced with:** Claude (Anthropic), August 2026, using live web search across roughly 200 tool calls via three parallel research agents — one on the commercial landscape, one on open-source agent and orchestration frameworks, one on open-source data, CRM and execution tooling. - **Method:** every factual claim was sourced from a primary or named source at time of writing and is linked in the Sources section. Where a claim came from a vendor with a commercial interest — competitor comparisons, vendor-authored landscape posts, self-reported benchmarks — that is flagged inline. > [!bug] Known limitations, stated honestly > - GitHub star counts were scraped and some look inflated. Treat them as order-of-magnitude only. Verify before quoting externally. > - Pricing changes constantly. Every figure is an August 2026 snapshot; several are already likely stale. > - A small number of items could not be fully verified and were marked ⚠️ in the original research. > - The recommended stack reflects an existing bias toward Anthropic tooling. ==That bias is exactly what Part B, question 2 exists to challenge.== **What this document is not:** a decision, a build plan, or a substitute for hands-on evaluation. It is a map of the territory as of one afternoon's research. > [!quote] > If you find something in Part C that is wrong, that finding is more valuable than the rest of the document. Flag it.